Add multi-provider AI configuration

[c-sonnier]

Add Multi-Provider AI Configuration

Summary

Adds per-account AI provider configuration, letting each account choose their own AI provider (Anthropic, OpenAI, Google Gemini, etc.) and model for the AI-powered features (error summaries, performance summaries, PR content generation, branch naming, code fix application). Replaces the hardcoded claude-haiku-4-5 / ANTHROPIC_API_KEY-only integration with a pluggable, DB-backed config layered over a unified RubyLLM interface.

Deployments that set ANTHROPIC_API_KEY (or any supported provider's ENV var) continue to work unchanged — accounts without an explicit config fall back to ENV automatically.

Motivation

• Accounts want to bring their own provider/keys (cost, data residency, model preference).
• AI services duplicated Anthropic HTTP wiring across four+ classes. A shared RubyLLM-based concern removes that duplication and makes adding providers a config change, not a code change.
• Single-provider lock-in blocked onboarding customers with existing OpenAI/Gemini contracts.

Screenshots

What's Included

New

• AiProviderConfig model (account_id, provider, model, encrypted api_key, active) with one active config per account enforced via activate!.
• Account-scoped CRUD at /account/settings/ai_provider_configs (create, update, destroy, activate, models) with AiProviderConfigsController.
• Settings UI — Stimulus controller (ai_provider_form_controller.js) with dynamic model list fetched per provider, search/filter, masked key input, loading and failure states.
• Shared concern AiProviderChat — one place where every AI service acquires a configured RubyLLM chat instance.
• Migration create_ai_provider_configs — additive only, with FK + index on account_id.
• Specs for model, request, factory, and updated service specs.

Changed

• AiSummaryService, AiPerformanceSummaryService, Github::BranchNameGenerator, Github::PrContentGenerator, Github::SimpleCodeFixApplier — all now take account: and delegate provider/key resolution to AiProviderChat.
• Account#ai_provider_config returns the persisted active config or an ENV-backed transient config if none exists.
• active_record_encryption.rb initializer now ENV-only, fails fast in production when keys missing (required since AiProviderConfig#api_key is encrypted).
• SUPER_ADMIN_WRITABLE_CONTROLLERS includes ai_provider_configs so super admins can manage provider configs while viewing another account.

Dependencies

• ruby_llm 1.14.1 (+ transitive gems in Gemfile.lock).

Backwards Compatibility

• No breaking changes for existing deployments.
• ENV-only setup (e.g. ANTHROPIC_API_KEY) continues to work via the Account#ai_provider_config fallback.
• AiSummaryJob#perform signature unchanged — in-flight jobs from prior code continue to succeed.
• Migration is additive; no data migration needed.
• No existing data is encrypted by the new encryption initializer (this is the first model to use encrypts), so ENV vars need to be set only before creating the first AiProviderConfig.

Deployment Notes

Before merging, production needs the three Active Record encryption ENV vars:

ACTIVE_RECORD_ENCRYPTION_PRIMARY_KEY
ACTIVE_RECORD_ENCRYPTION_DETERMINISTIC_KEY
ACTIVE_RECORD_ENCRYPTION_KEY_DERIVATION_SALT

Generate with bin/rails db:encryption:init. The initializer raises on boot in production if any are missing.

Testing

• RSpec: 44/44 target specs pass (model, request, services, account).
• Minitest: 3/3 remaining AI service tests pass (stale ai_summary_service_test.rb deleted — superseded by RSpec).
• GitHub service specs: 62/62 pass.
• Manually exercised the settings UI end-to-end: provider add/update/delete/activate, model dropdown fetch, unconfigured-state messaging, and AI summary generation against two providers.

Test Plan

• bundle exec rspec passes
• bundle exec rails test passes
• Fresh deploy with all three encryption ENV vars set — boots without error
• Fresh deploy without encryption ENV vars — raises on boot in production (expected)
• Existing deploy with only ANTHROPIC_API_KEY — AI summary generates without any DB config
• Create Anthropic config → generate AI summary → verify uses configured model
• Switch active config to OpenAI → generate AI summary → verify uses OpenAI
• Super admin viewing another account → can create/activate/delete configs